End-to-end deep learning for superoscillatory subtraction imaging

Yiping Lu, Qiuyu Ren, Zhigang Dai, Rui Yao, K Chen, Zhi Hong, Bin Fang, Fangzhou Shu, Shiyi Mei, Zhongwei JinarXiv 2025 · DOI · arXiv

TL;DR

Researchers have developed an end-to-end deep learning framework that merges super-oscillatory focusing with subtraction imaging to break the optical diffraction limit. By replacing manual parameter tuning with a jointly-optimized neural network, the system achieves sub-100-nm resolution while significantly reducing the artifacts and information loss typical of traditional label-free super-resolution methods.

What Problem Does This Paper Address?

In the world of optical microscopy, the "diffraction limit" is the traditional barrier—roughly 250 nm for visible light—that prevents us from seeing finer subwavelength details. While techniques like STED and STORM have shattered this limit, they often require fluorescent labels, which isn't always feasible for every sample.

Two promising label-free alternatives are Super-Oscillatory (SO) imaging (using interference to create tiny focal spots) and subtraction imaging (subtracting a "donut-shaped" dark-field image from a bright-field image to sharpen the effective point spread function). However, combining them is tricky. Traditionally, this requires a two-step acquisition process and the manual, subjective tuning of a subtraction coefficient ($\gamma$). If $\gamma$ is off, you get "negative dips" in your data, leading to lost information and halo artifacts. This paper aims to automate and optimize this marriage of techniques using deep learning.

Methods

The authors utilize a Vectorial Debye Integral Neural Network (VDINN). This is a physics-informed architecture designed to simulate and optimize light focusing in high-numerical aperture (NA) systems.

Key technical innovations include:

  • End-to-End Integration: Instead of treating wavefront modulation and image subtraction as separate steps, the network optimizes them simultaneously. It "launches" a plane wave and a vortex beam through the same optimized mask within the digital pipeline.
  • Learnable Subtraction: The subtraction coefficient ($\gamma$), previously a "best guess" hyperparameter, is transformed into a learnable parameter that the network optimizes during training.
  • Jointly-Optimized PSF: The network's loss function (Mean Squared Error) is designed to suppress high-energy sidelobes and minimize the deep negative values that usually plague subtraction imaging.
  • Confocal Integration: The system incorporates a virtual confocal pinhole (0.5 Airy units) to further clean up the signal and remove out-of-focus light.

Datasets

The study relies on high-fidelity physical simulations rather than standard image datasets like ImageNet:

  • Ground Truth Tensors: The network is trained against desired focal field distributions represented as (1024, 1024) tensors.
  • Periodic Stripe Patterns: To evaluate resolution, the authors used simulated periodic objects with features at 98 nm, 100 nm, and 105 nm.
  • Baseline Comparisons: Results were compared against standard SO focusing and a "simple" (non-optimized) combination of SO and subtraction imaging.

Key Insights and Ideas

  • The "Aha" Moment on Information Loss: The researchers found that simply adding a vortex phase to an existing SO mask (the "simple" approach) failed at high resolutions because it created massive negative dips in the Point Spread Function. These dips act like "black holes" for data during post-processing.
  • Neural Networks as Physical Compensators: By training the network end-to-end, the VDINN learned to adjust the amplitude and shape of the bright-field and dark-field beams so they matched more effectively. This resulted in a much higher learned $\gamma$ (0.4719 vs. the manual 0.006), leading to a sharper, cleaner focus.
  • Sidelobe Suppression: The intrinsic nature of the MSE loss function in the VDINN naturally "pushes down" the unwanted rings of light (sidelobes) that usually surround super-oscillatory spots, which is a major hurdle in practical SO microscopy.

Key Results

  • FWHM Reduction: The optimized Super-Oscillatory Subtraction Imaging (SOSI) achieved a Full Width at Half Maximum (FWHM) of 0.284 $\lambda_m$ (~99 nm), compared to 0.315 $\lambda_m$ for standard SO focusing.
  • Resolution Milestone: In imaging tests, the system successfully resolved periodic structures down to 98 nm using a 532 nm laser source.
  • Artifact Reduction: The relative depth of negative dips was reduced from 4.5% in the simple scheme to just 3% in the end-to-end scheme, nearly eliminating the "halo" artifacts that typically blur subtraction-based images.

Limitations and Future Work

While the results are impressive, the authors note a few areas for growth:

  • Beam Selection: The study used a first-order vortex beam for the dark-field component because it is straightforward. However, other structured light beams (like azimuthally polarized beams) might offer even better performance.
  • Simulation vs. Reality: The paper is primarily a simulation-based proof-of-concept. While the VDINN model is physically rigorous, real-world experimental noise and optical imperfections will provide the next big test.
  • Generalization: The framework is currently tuned for specific subwavelength tasks; extending it to handle more complex, non-periodic biological samples is a logical next step.

Summary updated

Abstract

Breaking the diffraction limit in optical imaging is crucial for resolving subwavelength details in a wide range of applications, where superoscillatory imaging and subtraction imaging are two common strategies for surpassing conventional resolution limits. We propose an end-to-end deep learning framework that integrates superoscillatory focusing and subtraction imaging into a single jointly-optimized vectorial Debye integral neural network pipeline, eliminating the traditional two-step acquisition and manual weighting process. With this end-to-end neural network, we further improve the focusing capability of the system to the sub-100-nm regime, enabling deep-subwavelength imaging resolution.
BibTeX
@misc{lu2025end,
  author = {Yiping Lu and Qiuyu Ren and Zhigang Dai and Rui Yao and K Chen and Zhi Hong and Bin Fang and Fangzhou Shu and Shiyi Mei and Zhongwei Jin},
  title = {{End-to-end deep learning for superoscillatory subtraction imaging}},
  year = {2025},
  eprint = {2511.16115},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2511.16115},
}
Figure